requirements.txt

# --- Core PyTorch and CUDA ---
# Ensure this matches your CUDA 12.1. Use the official PyTorch website to get the correct command.
torch>=2.1.0 --index-url https://download.pytorch.org/whl/cu121 # Or newer (e.g., 2.2, 2.3 if stable)
# torchvision torchaudio # Often bundled, include if you have direct uses for them

# --- Hugging Face Ecosystem (Essential for your LLM tasks) ---
transformers>=4.38.0 # Keep very up-to-date for latest features, models, and optimizations like sdpa
accelerate>=0.27.0   # Crucial for multi-GPU (device_map="auto") and other training/inference utilities
datasets             # For loading and processing datasets if you fine-tune or evaluate
huggingface_hub      # For interacting with the Hugging Face Hub (downloading models, etc.)
safetensors          # Often a dependency for secure model weight handling

# --- Quantization & Performance ---
bitsandbytes>=0.42.0 # For 4-bit/8-bit quantization (NF4, etc.). Ensure compatibility with your CUDA version.
auto-gptq>=0.7.0     # For running GPTQ quantized models. Keep up-to-date for new features/kernels.
                     # Consider a specific build if available for your CUDA version if generic pip install has issues.
                     # NOTE: Use python version 3.10 (<=3.12), and load module (command shown here): module load cuda11.1
# optimum>=1.17.0    # For advanced model optimization, export to ONNX, and some specific hardware acceleration (e.com/NVIDIA/FasterTransformer).
                     # Also, some AutoGPTQ backends (like exllamav2) might leverage optimum.
peft>=0.9.0          # Parameter-Efficient Fine-Tuning (LoRA, etc.), if you plan to fine-tune
# flash-attn          # Optional: For FlashAttention v1 on V100s. Install with --no-build-isolation.
                     # Test carefully against 'sdpa'. Might not be necessary.
# xformers           # Optional: Another library for optimized attention mechanisms.
                     # Was more critical before PyTorch 2.0's 'sdpa'. Can sometimes still offer benefits.

# --- GGUF Model Execution (Alternative to Hugging Face Transformers for some models) ---
ctransformers        # For running GGUF quantized models (often used with llama.cpp).
                     # Can be CPU or GPU accelerated. Ensure you install with CUDA support if using on GPU:
                     # CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=70" pip install ctransformers --no-binary ctransformers
                     # (70 is for Volta - V100. Adjust if necessary)

# --- NLP Evaluation & Utilities ---
nltk                 # Natural Language Toolkit for text processing tasks
rouge-score          # For ROUGE evaluation metric (summarization, etc.)
bert-score           # For BERTScore evaluation metric (semantic similarity)
sentence-transformers # For generating sentence embeddings
tqdm                 # For progress bars, very useful
python-Levenshtein

# --- Data Handling ---
pandas               # For data manipulation and analysis (DataFrames)
numpy                # Fundamental package for numerical computation

# --- API Access ---
openai               # If you are using OpenAI APIs

# --- Application Specific (Chess related) ---
python-chess         # For chess logic in Python
stockfish            # Python wrapper for the Stockfish chess engine

# --- Potentially Missing (Consider if relevant) ---
# tensorboard        # For logging and visualizing training runs (if you get into fine-tuning)
# wandb              # Weights & Biases: Another popular tool for experiment tracking & visualization

# --- General Good Practices ---
# Consider using a virtual environment (venv, conda) to manage these dependencies.
# Regularly update packages, but test after updates as breaking changes can occur.
# You can pin versions (e.g., transformers==4.38.1) once you have a stable setup.

tensorboard